Hybrid AI-Physics Framework for Lost Circulation Zone Diagnosis: A Data-Driven Iterative Approach Integrating Drilling-Logging Big Data with Theoretical Models

Summary This paper develops a hybrid AI-physics framework designed to enable real-time, cost-effective diagnosis of loss zones in complex formations. The framework integrates drilling and surface logging big data with physics-based models to address two critical industry challenges: (1) the high uncertainty of manual theoretical calculations and (2) the delays and costs of post-event instrument-based measurements. By using real-time data, it rapidly determines loss zone characteristics (depth, channel type, and fracture dimensions), enhancing treatment success rates and holding the potential to reduce operational costs. A Sichuan Basin database was constructed from over 100 wells using an event-based approach, capturing time-series data around lost circulation events. Theoretical models for determining loss zone depth, identifying lost circulation channel types, and estimating fracture width were evaluated through statistical error analysis. AI optimization methods tailored to each theoretical model were selected, and corresponding AI-physics hybrid models were developed independently, followed by training and validating these models on a database randomly split with 80% allocated for training and 20% for validation. These hybrid models were embedded into the Engineering Intelligent Support Center (EISC) with real-time drilling and surface logging parameters to establish loss zone diagnostic framework. A machine learning-enhanced drilling hydraulics model was developed to calculate the loss zone depth using real-time drilling and surface logging parameters, achieving a test set mean error of 4.06%-4.87%, representing a 10%-48% reduction compared to theoretical models. A convolutional neural network-based intelligent classification model was established to identify lost circulation channel types, leveraging geometric features of loss rate vs. time and mud pit volume vs. time curves to distinguish natural fractures, induced fractures, microfractures/high-permeability pores, and medium-to-large fractures/caverns, with a test set accuracy exceeding 94%. An ensemble learning model integrating AdaBoost with solid- and liquid-phase lost circulation dynamics was constructed to invert fracture width, utilizing drilling fluid or lost circulation material slurry flow data, reducing test set mean errors to 12.07%-42.73%, an improvement of up to 73.46% over theoretical methods. Field trials across over 20 wells in the Sichuan Basin demonstrated that, following post-lost circulation events, engineers could access real-time drilling and surface logging data via the field-deployed Framework connected to the EISC, rapidly determining the loss zone depth, channel type, and fracture width to inform lost circulation control strategies. The framework achieved a field-validated diagnostic accuracy exceeding 85%, with ongoing updates to lost circulation data expected to further enhance precision. This work introduces the first EISC-integrated hybrid AI-physics framework for loss zone diagnosis. It bridges the gap between data-driven AI and physics-based models, enabling rapid, cost-effective diagnosis with acceptable accuracy and without additional downhole instrumentation. The framework offers engineers a practically validated tool for decision-making in lost circulation treatment, balancing theoretical rigor with AI adaptability.